fix(tabular): union passthrough CSV schema and keep assistant text visible

[Paul Lizer (paullizer)]

Summary

Two customer-reported failures in chat CSV/tabular artifact creation.

1. Passthrough CSV exports failed with a schema mismatch

Chat and workflow CSV artifacts stage already-final rows with passthrough_input_rows=True. _build_passthrough_batch_results fed those rows to _normalize_generated_batch_entries as both source and "generated" entries, so the output schema was derived from row 1 only and every later row had to match it exactly.

extract_authorized_function_result_rows merges rows from multiple action results into a single set, so field sets legitimately differ. Real failure:

Generated output schema mismatch at row 2; missing=['facts'];
unexpected=['application', 'message_metadata', 'runtime', 'schema_version']

It was then classified model_validation retryable and burned all three automatic retries in ~20 ms each -- no model runs on the passthrough path at all -- before failing the run permanently.

Fix: queue_tabular_generated_output_run unions column names across every staged passthrough batch and pins output_schema / public_output_schema / internal_checkpoint_schema up front. _build_passthrough_batch_results pads sparse rows to that schema, with a per-batch union fallback for runs queued before this change. Model-generated runs still defer to batch-1 schema discovery.

This matches build_assistant_table_csv and _build_generated_output_csv, which already unioned columns on the synchronous path.

2. The CSV artifact card hid the assistant response

hideCompletedGeneratedArtifactHandoff() adds d-none to .message-text and .message-footer when an artifact sets suppress_assistant_text. That flag exists for JSON/XML, whose payload is deliberately withheld from the streamed text (suppress_streamed_file_payload = requested_streamed_file_format in {'json', 'xml'}) and replaced with a "Generating the X file..." placeholder.

CSV was wrongly in the same set, so a full CSV narrative -- including inline charts -- was hidden the moment the card hydrated.

Fix: added ASSISTANT_TEXT_SUPPRESSING_FORMATS = {'json', 'xml'} and reused it for background structured exports. Markdown analysis artifacts keep suppress_assistant_text=True; that one is a genuine handoff.

Changes

• application/single_app/functions_tabular_generated_exports.py -- union passthrough schema at queue time, _extend_passthrough_output_field_names, _align_passthrough_rows_to_output_fields, format-aware suppression
• application/single_app/functions_generated_file_exports.py -- ASSISTANT_TEXT_SUPPRESSING_FORMATS
• application/single_app/config.py -- version 0.260.005
• functional_tests/test_tabular_passthrough_heterogeneous_rows.py -- new
• functional_tests/test_assistant_table_csv_artifact.py -- assistant-text preservation coverage

Validation

New test_tabular_passthrough_heterogeneous_rows.py (4/4) exercises the real queue_tabular_generated_output_run and _build_passthrough_batch_results via the AST harness, with only Cosmos/blob/telemetry stubbed.

Suite	Result
test_tabular_passthrough_heterogeneous_rows.py	4/4
test_assistant_table_csv_artifact.py	36/36
test_generated_json_xml_exports.py	7/7
test_tabular_background_generated_exports.py	8/8
test_tabular_queue_run_output_schema_end_to_end.py	3/3
test_tabular_row_orchestration_scale.py	15/15
test_tabular_phase5_artifact_set_lifecycle.py	4/4
test_tabular_combined_artifact_set_download_visibility.py	3/3
test_tabular_durable_artifact_lifecycle_recovery.py	5/5
test_tabular_search_analyze_artifact_matrix.py	1/1
test_tabular_shared_request_planner.py	8/8
ui_tests/test_chat_background_generated_export_status.py	7 passed, 4 skipped (no SIMPLECHAT_UI_BASE_URL)

py_compile clean, editor diagnostics clean, Windows-safe git diff --check clean.

Notes

• Runs that already failed with the schema mismatch cannot self-heal; the request must be re-issued after deploy.
• Release notes not updated in this PR -- happy to add an entry if preferred.